Ooh, I will be giving this a go!

I mean, for anything where you’re willing to trust the container provider not to push breaking changes, you can just run Watchtower and have it automatically update. That’s how most of my stuff runs.

For the level of investment in and hype around this company? Yes, those enterprise sales are abysmal. When there are major news articles about their product every single week, they should be doing a lot better than that.

They have demonstrated zero ability at actually “hyperscale”. They have no path to getting those costs down. Their conversion rate from free to paid users is atrocious, and they’re already raising prices on their plans which is only going to worsen those conversion rates. Their costs to build future models are ballooning exponentially, and theres a decent chance that at some point Microsoft will get sick of subsidizing their compute costs.

Is it possible that they could be successful? Yes. But a lottery ticket would probably be a sounder investment.

For the record, OpenAI themselves are telling those VCs that they should think of their investment more as a “donation” with no expectation of future profit. Absolutely oozing confidence there.

There’s no good answer to that because it depends entirely on what you’re running. In a magical world where every open source project always uses the latest versions of everything while also maintaining extensive backwards compatibility, it would never be a problem. And I would finally get my unicorn and rainbows would cure cancer.

In practice, containers provide a layer of insurance that it just makes no sense to go without.

From a nerdy perspective, LLMs are actually very cool. The problem is that they’re grotesquely inefficient. That means that, practically speaking, whatever cool use you come up with for them has to work in one of two ways; either a user runs it themselves, typically very slowly or on a pretty powerful computer, or it runs as a cloud service, in which case that cloud service has to figure out how to be profitable.

Right now we’re not being exposed to the true cost of these models. Everyone is in the “give it out cheap / free to get people hooked” stage. Once the bill comes due, very few of these projects will be cool enough to justify their costs.

Like, would you pay $50/month for NotebookLM? However good it is, I’m guessing it’s probably not that good. Maybe it is. Maybe that’s a reasonable price to you. It’s probably not a reasonable price to enough people to sustain serious development on it.

That’s the problem. LLMs are cool, but mostly in a “Hey this is kind of neat” way. They do things that are useful, but not essential, but they do so at an operating cost that only works for things that are essential. You can’t run them on fun money, but you can’t make a convincing case for selling them at serious money.

How is that not worth a boatload of money?

Because they spend $2.35 billion in operating costs for every $1 billion in revenue (gross, not net).

OpenAI loses money at an absolutely staggering rate, and every indication, even their own openly stated predictions, are that those costs will only increase.

And keep in mind, right now OpenAI gets a lot of their investment in the form of compute credits from Microsoft, which they get to spend at a massively discounted rate. That means that if they were actually buying their Azure time at market value they’d be closer to spending something like $5bn to make $1bn.

Again, I really need to be clear here, I’m not saying “to make 1 billion in profit.” I’m saying “revenue”. They lose money every time someone uses their services. The more users they have, the more their losses grow. Even paid users cost them more money than they pay in most cases.

This is like a store that buys products at $10 and sells them at $4. It is the most insanely unprofitable business plan imaginable.

And it’s not getting better. Conversions to paid plans are at about 3%. Their enterprise sales are abysmal. Training costs are increasing exponentially with each new generation of models. Attempts to make their models more compute efficient have so far failed utterly.

OpenAI’s path to profitability is basically “Invent true AGI.” It’s a wild fantasy with zero basis in reality that investors are shovelling money into because investors will shovel money into anything that promises infinite growth.

Are sales bad?

Of AI products? By all available metrics, yes, sales for AI driven products are atrocious.

Even the biggest name in AI is desperately unprofitable. OpenAI has only succeeded in converting 3% of their free users to paid users. To put that on perspective, 40% of regular Spotify users are on premium plans.

And those paid plans don’t even cover what it costs to run the service for those users. Currently OpenAI are intending to double their subscription costs over the next five years, and that still won’t be enough to make their service profitable. And that’s assuming that they don’t lose subscribers over those increased costs. When their conversion rate at their current price is only 3%, there’s not exactly an obvious appetite to pay more for the same thing.

And that’s the headline name. The key driver of the industry. And the numbers are just as bad everywhere else you look, either terrible, or deliberately obfuscated (remember, these companies sank billions of capex into this; if sales were good they’d be talking very openly and clearly about just how good they are).

That’s not what the article says.

They’re arguing that AI hype is being used as a way of driving customers towards cloud infrastructure over on-prem. Once a company makes that choice, it’s very hard to get them to go back.

They’re not saying that AI infrastructure specifically can be repurposed, just that in general these companies will get some extra cloud business out of the situation.

AI infrastructure is highly specialized, and much like ASICs for the blockchain nonsense, will be somewhere between “very hard” and “impossible” to repurpose.

It’s a bubble because OpenAI spend $2.35 for every $1.00 they make. Yes, you’re mathing right, that is a net loss.

It’s a bubble because all of the big players in AI development agree that future models will cost exponentially more money to train, for incremental gains. That means there is no path forward that doesn’t intensely amplify the unprofitability of an already deeply unprofitable industry.

It’s a bubble because newer models with better capabilities only cost more and more to run.

It’s a bubble because as far as anyone knows there will never be a solution to the hallucination problem.

It’s a bubble because despite investments treating it as a trillion dollar industry, no one has yet figured out a trillion dollar problem that AI can solve.

You’re trying on a new top of the line VR headset and saying “Wow, this is incredible, how can anyone say this is a bubble?” Its not about how cool the tech is in isolation, it’s about its potential to effect widespread change. Facebook went in hard on VR, imagining a future where everyone worked from home while wearing VR headsets. But what they got was an expensive toy that only had niche uses.

AI performs do well on certain coding tasks because a lot of the individual problems that make up a particular piece of software have already been solved. It’s standard practice to design programs as individual units, each of which performs the smallest task possible, and which can then be assembled to complete more complex tasks. This fits very well into the LLM model of assembling pieces into their most likely expected configurations. But it cannot create truly novel code, except by a kind of trial and error mutation process. It cannot problem solve. It cannot identify a users needs and come up with ideal solutions to them. It cannot innovate.

This means that, at best, genAI in the software world becomes a tool for producing individual code elements, guided and shepherded by experienced programmers. It does not replace the software industry, merely augments it, and it does so at a cost that many companies simply may not feel is worth paying.

And that’s its best case scenario. In every other industry AI has been a spectacular failure. But it’s being invested in as if it will be a technological reckoning for every form of intellectual labour on earth. That is the absolute definition of a bubble.

The answer is that it’s all about “growth”. The fetishization of shareholders has reached its logical conclusion, and now the only value companies have is in growth. Not profit, not stability, not a reliable customer base or a product people will want. The only thing that matters is if you can make your share price increase faster than the interest on a bond (which is pretty high right now).

To make share price go up like that, you have to do one of two things; show that you’re bringing in new customers, or show that you can make your existing customers pay more.

For the big tech companies, there are no new customers left. The whole planet is online. Everyone who wants to use their services is using their services. So they have to find new things to sell instead.

And that’s what “AI” looked like it was going to be. LLMs burst onto the scene promising to replace entire industries, entire workforces. Huge new opportunities for growth. Lacking anything else, big tech went in HARD on this, throwing untold billions at partnerships, acquisitions, and infrastructure.

And now they have to show investors that it was worth it. Which means they have to produce metrics that show people are paying for, or might pay for, AI flavoured products. That’s why they’re shoving it into everything they can. If they put AI in notepad then they can claim that every time you open notepad you’re “engaging” with one of their AI products. If they put Recall on your PC, every Windows user becomes an AI user. Google can now claim that every search is an AI interaction because of the bad summary that no one reads. The point is to show “engagement”, “interest”, which they can then use to promise that down the line huge piles of money will fall out of this pinata.

The hype is all artificial. They need to hype these products so that people will pay attention to them, because they need to keep pretending that their massive investments got them in on the ground floor of a trillion dollar industry, and weren’t just them setting huge piles of money on fire.

Personally, I always like to use containers when possible. Keep in mind that unlike virts, containers have very minimal overhead. So there really is no practical cost to using them, and they provide better (though not perfect) security and some amount of sandboxing for every application.

Containers mean that you never have to worry about whether your VM is running the right versions of certain libraries. You never have to be afraid of breaking your setup by running a software update. They’re simpler, more robust and more reliable. There are almost no practical arguments against using them.

And if you’re running multiple services the advantages only multiply because now you no longer have to worry about running a bespoke environment for each service just to avoid conflicts.

Ombi

Yeah, my own experience of switching to containers was certainly frustrating at first because I was so used to doing things the old way, but once it clicked I couldn’t believe how much easier it made things. I used to block out several days for the trial and error it would take getting some new service to work properly. Now I’ll have stuff up and running in 5 minutes. It’s insane.

While I understand the frustration of feeling like you’re being forced to adopt a particular process rather than being allowed to control your setup the way you see fit, the rapid proliferation of containers happened because they really do offer astonishing advantages over traditional methods of software development.

I’ll add here that the “docker top” command allows you to easily see what kind of resources your containers are using.

If you prefer a UI, Dozzle runs as a container, is super lightweight, requires basically no setup, and makes it very easy to see your docker resource usage.

Correct on both counts, although it is possible to set limits that will prevent a single container using all your system’s resources.

Your must get, like, 10fps on a rig like that.

Again, what you’re not clocking here is that it will be a very, very long time before we have sufficient quantum compute time available to engage in large scale decryption. Even just getting to the point where they can decrypt all newly generated messages will be a long time. By that point you’d have decades of historical messages to did through.

Barring some wild, out of nowhere leap forward in the feasibility, scalability and affordability of the tech, you’ll be dead by the time the NSA gets around to reading your old messages.

For the record, these aren’t consumer GPUs. They’re a completely different beast, designed specifically for running transformer models and costing up to $70,000 each.

In the case of quantum computing, there is a real meaning to it (in really vague terms, its computing using the suoerposition of quantum states to collapse extraordinarily complex problems down to a single answer). The problem rather is that right now companies are eagerly hyping this tech as being “just around the corner” when it’s nothing of the sort (unless a bunch of massive breakthroughs suddenly turn up).